From bc5b7a5fb5020f0611f4632bffc81a20041a63bb Mon Sep 17 00:00:00 2001
From: Wang Kunpeng <1289706727@qq.com>
Date: Fri, 26 Dec 2025 21:37:28 +0800
Subject: [PATCH] [bugfix] Fix MHA model runtime error in aclgraph mode (#5397)

### What this PR does / why we need it?
Currently, MHA models (eg: minicpm-2b, Baichuan-7b) will encounter
errors when running in piecewise graph mode, with error messages similar
to:
```
(E89999):  When layout is TND and PA not enabled, keyT(8) and valueT(8) must be equal to the last element of actualSeqenceLengthKV(5)[FUNC:CheckInputShapeWhenLayoutIsTND][FILE:prompt_flash_attention_tiling.cpp][LINE:3618]
```
The error occurs because the qkv in the Prefill stage is also padded,
causing the shape to be inconsistent with actual_seq_lengths.
Add unpadding logic for kv.

- vLLM version: release/v0.13.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/254f6b986720c92ddf97fbb1a6a6465da8e87e29

Signed-off-by: Wang Kunpeng <1289706727@qq.com>
---
 .github/workflows/_e2e_test.yaml      |  3 +--
 tests/e2e/singlecard/test_models.py   | 14 ++++++++++++--
 vllm_ascend/attention/attention_v1.py |  3 +++
 3 files changed, 16 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/_e2e_test.yaml b/.github/workflows/_e2e_test.yaml
index 56df80ad..6d408f1c 100644
--- a/.github/workflows/_e2e_test.yaml
+++ b/.github/workflows/_e2e_test.yaml
@@ -115,8 +115,7 @@ jobs:
           pytest -sv --durations=0 tests/e2e/singlecard/test_sampler.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_vlm.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_xlite.py
-          # TODO: revert me when minicpm is fixed
-          # pytest -sv --durations=0 tests/e2e/singlecard/test_models.py
+          pytest -sv --durations=0 tests/e2e/singlecard/test_models.py
           pytest -sv --durations=0 tests/e2e/singlecard/pooling/
           pytest -sv --durations=0 tests/e2e/singlecard/compile/test_norm_quant_fusion.py
           pytest -sv --durations=0 tests/e2e/singlecard/test_multistream_overlap_shared_expert.py
diff --git a/tests/e2e/singlecard/test_models.py b/tests/e2e/singlecard/test_models.py
index 20fe6f77..8068eda5 100644
--- a/tests/e2e/singlecard/test_models.py
+++ b/tests/e2e/singlecard/test_models.py
@@ -19,18 +19,28 @@
 
 import os
 
+import pytest
+from modelscope import snapshot_download  # type: ignore
+
 from tests.e2e.conftest import VllmRunner
 
 os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 
+# Note: MiniCPM-2B is a MHA model, MiniCPM4-0.5B is a GQA model
+MINICPM_MODELS = [
+    "openbmb/MiniCPM-2B-sft-bf16",
+    "OpenBMB/MiniCPM4-0.5B",
+]
 
-def test_minicpm_2b() -> None:
+
+@pytest.mark.parametrize("model", MINICPM_MODELS)
+def test_minicpm(model) -> None:
     example_prompts = [
         "Hello, my name is",
     ]
     max_tokens = 5
 
-    with VllmRunner("openbmb/MiniCPM-2B-sft-bf16",
+    with VllmRunner(snapshot_download(model),
                     max_model_len=512,
                     gpu_memory_utilization=0.7) as runner:
         runner.generate_greedy(example_prompts, max_tokens)
diff --git a/vllm_ascend/attention/attention_v1.py b/vllm_ascend/attention/attention_v1.py
index 001d58fb..0aef154a 100644
--- a/vllm_ascend/attention/attention_v1.py
+++ b/vllm_ascend/attention/attention_v1.py
@@ -532,6 +532,9 @@ class AscendAttentionBackendImpl(AttentionImpl):
             = self._get_fia_params(key, value, attn_metadata)
         num_tokens = attn_metadata.actual_seq_lengths_q[-1]
         query = query[:num_tokens]
+        if attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
+            key = key[:num_tokens]
+            value = value[:num_tokens]
         # Get workspace from cache or calculate it if not present.
         attn_output, _ = torch_npu.npu_fused_infer_attention_score(
             query=query,